A deep learning method for delineating early gastric cancer resection margin under chromoendoscopy and white light endoscopy

Ping An,Dongmei Yang,Jing Wang,Lianlian Wu,Jie Zhou,Zhi Zeng,Xu Huang,Yong Xiao,Shan Hu,Yiyun Chen,Fang Yao,Mingwen Guo,Qi Wu,Yanning Yang,Honggang Yu

DOI: https://doi.org/10.1007/s10120-020-01071-7

Gastric Cancer

Abstract:Background: Accurate delineation of cancer margins is critical for endoscopic curative resection. This study aimed to train and validate real-time fully convolutional networks for delineating the resection margin of early gastric cancer (EGC) under indigo carmine chromoendoscopy (CE) or white light endoscopy (WLE), and evaluated its performance and that of magnifying endoscopy with narrow-band imaging (ME-NBI). Methods: We collected CE and WLE images of EGC lesions to train fully convolutional networks ENDOANGEL. ENDOANGEL was tested both on stationary images and endoscopic submucosal dissection (ESD) videos. The accuracy and reliability of ENDOANGEL and NBI-dependent delineation were further evaluated by a novel endoscopy-pathology point-to-point marking. Results: ENDOANGEL had an accuracy of 85.7% in the CE images and 88.9% in the WLE images under an overlap ratio threshold of 0.60 in comparison with the manual markers labeled by the experts. In the ESD videos, the resection margins predicted by ENDOANGEL covered all areas of high-grade intraepithelial neoplasia and cancers. The minimum distance between the margins predicted by ENDOANGEL and the histological cancer boundary was 3.44 ± 1.45 mm which outperformed the resection margin based on ME-NBI. Conclusions: ENDOANGEL has the potential to assist endoscopists in delineating the resection extent of EGC under CE or WLE during ESD.